针对日径流序列不平稳性与波动性的特点,尝试将模态分解和深度学习组合预测方法引入至日径流预报.首先,采用完整集合经验模态分解方法对日径流时间序列进行分解,得到不同频率成分的模态分量;其次,为不同的模态分量构建长短期记忆神经网络日径流预测模型,并采用网格搜索参数寻优算法优化预报模型的超参数;最后将各模型预报结果进行模态重组获得日径流预报结果.以宜昌水文站日径流预报为研究实例,所构建的组合模型较单一长短期记忆神经网络的和分别降低了65.02%、58.35%与2.88%,且运用完整集合经验模态分解方法的分解效果优于传统的模态分解方法,为短时间尺度下的非线性和非平稳的日径流预报提供了新的方法和参考.
A combination of modal decomposition and deep learning forecasting methods was introduced to daily runoff forecasting to address the characteristics of unstable and volatile daily runoff series.Firstly,the complete ensemble empirical modal decomposition method was used to decompose the daily runoff time series,so as to obtain the modal components of different frequency components.Secondly,the daily runoff forecasting model was constructed for different modal components based on the long short-term memory neural network(LSTM),and the hyperparameters of the forecasting model were optimized using the grid search parametric optimization algorithm.Finally,the forecasting results of each model were modally reconstructed to obtain daily runoff forecasting results.The daily runoff forecasting of the Yichang hydrological station was taken as an example.Compared with the single LSTM,the RMSE,MAE,and MAPE of the proposed combination model were reduced by 65.02%,58.35%,and 2.88%,respectively.The decomposition effect of the complete ensemble empirical mode decomposition was better than that of the traditional modal decomposition method,which provided a new method and reference for nonlinear and non-stable daily runoff forecasting in a short time scale.